News for hackers. Stuff that matters.
Re:Show HN: A 6M-token movable window on a single 46GB GPU by SwellJoe on Tuesday July 28, 2026 @02:38AM (#49080191)
So, you don't know what it is, either.
Re:Show HN: A 6M-token movable window on a single 46GB GPU by Wetime on Tuesday July 28, 2026 @02:42AM (#49080219)
We handeld llm like a human brain we decopelled knowledge from the memory and build a memory layer that makes redoing things free and fast, so the llm can once it learned something solves it for free the next time
Re:Show HN: A 6M-token movable window on a single 46GB GPU by saidnooneever on Tuesday July 28, 2026 @03:28AM (#49080539)
this sounds like ur explaining caching
Re:Show HN: A 6M-token movable window on a single 46GB GPU by SwellJoe on Tuesday July 28, 2026 @10:44AM (#49084748)
Sounds like they're explaining magic. Because current models cannot learn and they cannot remember.
Re:Show HN: A 6M-token movable window on a single 46GB GPU by saidnooneever on Tuesday July 28, 2026 @10:47AM (#49084807)
there is RL for LLMs which actually changes the weights but its more specialization than learning and wont counter the probabalistic nature of the thing
Re:Show HN: A 6M-token movable window on a single 46GB GPU by SwellJoe on Tuesday July 28, 2026 @11:45AM (#49085639)
Training is not something you can just bolt on, and it generally requires even larger hardware than inference for a given model, and a huge amount of time. If you're aiming for "free", a OP claims, RL aint it.